Fix LayerNorm crash when model.half() is used

[OWU-4f5755]

LayerNorm.forward() casts the input to fp32 but doesn't cast its own weight/bias, so calling model.half() before transcription causes:

RuntimeError: expected scalar type Float but found Half

Linear and Conv1d in the same file already guard against this by casting their weights to match the input dtype. This PR does the same for LayerNorm.

# Before — weight/bias stay fp16 after model.half()
class LayerNorm(nn.LayerNorm):
    def forward(self, x: Tensor) -> Tensor:
        return super().forward(x.float()).type(x.dtype)

# After — explicitly cast weight/bias to fp32
class LayerNorm(nn.LayerNorm):
    def forward(self, x: Tensor) -> Tensor:
        return F.layer_norm(
            x.float(),
            self.normalized_shape,
            self.weight.float() if self.weight is not None else None,
            self.bias.float() if self.bias is not None else None,
            self.eps,
        ).type(x.dtype)

Observed no overhead in the normal case (.float() on an fp32 tensor is a no-op). Tested with model.half() + fp16=True and the standard path — both work.